Presenting DR34MSC4PE's WAN LoRA:
DR34MJOB - Single and Double Blowjobs (POV+)! It, as a nice surprise, is excellent at handjobs too!
Please remember to like and share - if you like our work, give us a shout below!
===========================================================================
9/20 - Come and get it... WAN 2.2 I2V updates are here!
Note: You need BOTH models to produce good output. If you don't like what you see - adjust the CFG!! We hope you enjoy. =)
===========================================================================
Sample Prompts:
d0ubl3_bj, A close-up view of two women performing a double blowjob on a man's erect penis. The woman on the left has light brown skin blonde slutty pigtail hair. Her nails are manicured with a glittery design as she grasps the base of the penis. The other woman has light tan-lined skin and skin black hair with bows and her eyes are closed. Both women are focused on the man's penis which is circumcised and erect. They both lick and run their lips along the shaft of the man's penis. The background includes a prison cell and bed, their orange jumpers cast aside their nude bodies on the floor. The lighting is bright and even highlighting the details of the subjects' skin and hair. The camera angle is a low close-up shot focusing on the women's faces and the POV 1penis they are pleasuring
bl0wj0b, A close-up view of a blonde woman with fair skin and a slim physique wearing a pink tank top. She is positioned on her knees leaning over a man who is lying on his back. The man has short brown hair and is dressed in a black suit jacket over a light blue shirt. The woman's mouth is open as she licks the head of the man's erect penis which is circumcised. Her eyes are focused intently on the man as she slowly strokes the penis with her right hand. Her left hand rests on the man's right thigh. The background includes a yellow couch and a wooden floor. The lighting is warm and soft creating a cozy atmosphere. The camera angle is slightly above and to the side of the woman providing a clear view of her facial expression and the action.
===========================================================================
DR34MSC4PE is
c0ur4ge (trainer/qa/inference code) /
ERA5ER (captions/tooling code/data/qa)
Be sure to check out our FLUX model - C4PACITOR!
Like our work? Buy us a coffee: https://ko-fi.com/dr34msc4pe
Please generate responsibly!
Keywords/Triggers:
Single BJ: bl0wj0b
Double BJ: d0ub13_bj
Description
All the goodness of T2V, specially tuned for 720p+ resolutions. Please enjoy responsibly!
FAQ
Comments (72)
This lora is CRACKED. It's not even funny how good it is. Super high fidelity, always works, can do multiple things with no issue. Truly the pinnacle 🙏🙏😭
Glad you’re loving it! New concepts and stuff on the horizon! Stay tuned.
This works really really well c0ur4ge, thanks for creating it and sharing it! What strength do you recommend for T2V?
Thanks - we love to hear it! Honestly I was just testing with 1.0 but it's probably worth dialing it back a bit to see where it gives the best fidelty without making the movements too jerky/close to vanilla WAN.
Trying to get only handjobs but having about a 25% success rate the girls usually end up sucking anyway, any prompts that will up the success rate, very nice lora though tysm.
Late to the party, you have luck with this?
Perhaps try adding to the negative "she takes the penis into her mouth" or "she sucks on the penis".
And make sure positive has something along the lines of 'she slowly/gently strokes the/his penis with her left/right hand' (customize as you wish).
Honestly the best LORA ive found. Pretty much all the others always distort the original image (I2V). Thanks so much for this mate. Which ever way you trained this to keep image identify, keep doing it for the I2V's.
Thanks! Be sure to check out our other models and LoRAs. WAN will definitely be getting some more love soon if we don’t see others start to raise the bar but as always we want to lead by example. :)
We appreciate you sharing your feedback and creations!
she comin' thru she comin' thru she suckin dick she suckin dick
The lora is perfect but the only issue i'm facing is with the mouth, as soon as she takes the pp in her mouth the upper and lower lips both stretch so much like only the mouth part becomes hella stretchy weird looking making it look soo creepy ;-; any suggestions on what to do to fix this will be highly appreciated!
Can you share your sampler settings/step count and whatnot? It's generally going to be something in there. I could also see it being potentially related to the aspect ratio of the input image relative to the desired generation size.
Question to anybody. Is anyone using SwarmUI? I've been using it with some success but I'm wondering if I should bite the bullet and use Comfy. Thanks.
I'm using SwarmUI and it has Comfy built in as a backend. It is preferable to use Comfy for Wan
Man this is absolutely crazy good. The best LORA i had for WAN. I have no idea how you did that but face consistency is amazing, it almost always keeps the original face with i2v! Please do more LORAs! :D
any prompt recomendations for i2v? the quality is awesome but having a bit of trouble to generate single person bj's
Glad you enjoy it! There's very little secret to the quality of our work - we have fantastic datasets and lots of trial and error! Please check out our other models and share your creations!
Hi Marqs, any chance you could share the workflow? Got other Loras up and running that also provided a workflow but still too much of a beginner to develop a workflow myself! Appreciate it
shitcrazy
Hey c0ur4ge, ive been trying for days to get this lora to work well with the causvid one. I can somewhat get them to work together...but man the videos end up looking like simple gif's lol. Have you tested out causvid yet with this lora? Or has anyone have any suggestions? ...ive tried lots
Try lowering the strength of causvid and any other combined loras starting at 0.50. It definitely helped me
with more realistic motions and especially with retaining resemblances. Although I'm still struggling to get
hand movements at all for unknown reasons.
So, I'm a bit of a purist when it comes to combining LoRA. It's never a guarantee that they'll work together because ultimately they are like two delicious sauces being mixed together. Sometimes, they'll combine to more than the sum of their parts but at the end of the day, its more likely to be less-than-better than either alone. This is particularly true at high strengths, so as they suggest below - start with our lora at ~0.8 and gradully increase the others (preferably have the lower strength lora FIRST in the chain) and then the one driving the higher strength (and desired overall outcome/content) last. Good luck!
How do you get all this motion in your videos? When I describe explicit behavior my videos are almost static. Is it the workflow, or is it dependant on if it's 1.3B or 14B?
Maybe you are using a cache with high strength, such as TeaCache? That will eliminate motion.
Alternatively maybe your negative prompt is stuff like "fast motion" etc.
Causvid lora typically dulls motion, try disabling it and set sampler settings to 20 with cfg at 4.
Stoked for the update!
Us too! Training starting in the coming days!
Hi! Are you willing to share more of your training settings (I'm trying to get a gauge for where things get overtrained)? I'm curious about your training set size, learning rate, rank. I assume your not using gradient accumulation? Do you use training loss as an indicator? Any evaluation metrics?
Hey there - no magic here! We've got very little "secret sauce" aside from just trying to "do it all right" re: hyperparameters and it clearly shows in the final product.
- ~75-100 videos
- Rank 16/32/64 (depending on concept complexity, start with 16)
- Good 'ol AdamW
- LR: 5e-5
- ~100 steps of warmup
- Grad accumulation of 4 steps
When it comes to LORA/Finetuning video transformers, typically convergence on a concept (provided its not too broad like the entire kama sutra :P) for WAN I see happens between 150-250 epochs with a dataset of that size, depending on your other hyperparameters (batch size, etc). This obviously depends on the diversity and complexity of the concept you're trying to train. The efficacy (or complete failure) of your training will hinge on the quality of your videos and captions and your ability to stop and sample 20-50 videos every 30 epochs or so!
The best way to not over train is to "stop when it's done" and you cannot know that without sampling frequently throughout training! Good luck!
Hey, amazing work!
is it possible to insert dick on dickless image?
Flux Context dev might do it... with the right lora.
Somewhat! Prompt for it as thought it is coming out of frame or as POV in the beginning and then describe what “it does” and I’ve definitely seen it work.
Absolutely yes. Prompt the dude into existence and you're good to go
Good
after months of using it with hundreds of amazing generations: This Lora is the absolute best!
Thanks again for the support - we take time to refine, so we appreciate in kind. ;)
can you share some configs please
like it❤️🔥
I must be using this lora wrong. I've been trying to get a side-view of a kneeling woman giving a blowjob for hours. This lora consistently stretches the penis to a weird length and has the male bend over to bring his face into the scene before either kissing the woman or licking his own dick. WTH?
If the video stubbornly keeps making a new person come into frame or keeps shooting liquid from people's faces, mouths, or eyes, is that fixable?
I feel like it's the model because it's definitely not just this lora doing it, but maybe its an issue with many loras? Sometimes you can stack huge amounts of prompting to get it to stop, but by then the prompt is so big its probably ruining the video quality and even that doesn't work most of the time.
Is this an issue for others too? I think the most likely seeming issue is maybe it's because i don't make realistic videos and the loras are trained so hard on realistic videos that it just generates a new person since it can't figure out the drawing style. Some of these loras show non-realistic examples though, so that ends up just being more confusing :D
I did get some eye cum. And also a hilarious out-the-nose one, which kind of worked XD
But only when prompting for cumshots, and using other loras for cumshots alongside this lora. Haven't tried it without.
Yeah these videos are trained on realism - we may incorporate some more anime into the training set eventually but it could definitely be that you’re using words that are fairly well associated with realistic people since that’s how we train the Lora.
c0ur4ge ah, it makes sense now :) Thanks for the reply!
Hi, great works, can you please provide me perfect parameters ?
width.height,steps,cfg,sampler.
thanks very much
Any that you’d use for WAN my friend. We train on multiple resolutions.
IMPORTANT: INFORM ALL WAN USERS! I DISCOVERED HOW TO USE LORAS 'NSFW/SFW' IN WAN 2.2 T2V, GETTING THE FULL POTENTIAL OF "HIGH NOISE" AND "LOW NOISE." THE RESULTS WERE IMPRESSIVE!! I ran it on an RTX 3060 6GB, 32GB RAM, 480x480, length: 41, generation time: 5 minutes to generate each video. The quality is fantastic! For those of you with more powerful computers, the results will be even more incredible. I used the models: WAN 2.2 - Q4_1.gguf FOLLOW THE STEPS BELOW: 1: Don't use acceleration/enhancement LoRas (such as Lightx2v or other) on "high noise"; use only on "low noise" with strength 1. (I'm using this version of Lightx2v LoRa): lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16 2: Use 'NSFW/SFW' LoRas on both "high noise" and "low noise" models. On "high noise," use strength 2, and on "low noise," use half the strength, with strength 1. You can vary the strengths as you like, but always keep the "low noise" strength at half the strength of the "high noise." 3: Use these settings in KSamplers: "high noise": steps: 9 / cfg: 3.5 / euler/simple / start_at_step: 0 / end_at_step: 4 "low noise": steps: 9 / cfg: 1.0 / euler/simple / start_at_step: 4 / end_at_step: 10000 Note: To improve speed, I set the Windows system to "best performance," disable unnecessary background programs and applications, and lower the desktop resolution. This significantly speeds up video generation. I hope this helps.
Thank you, I could not do a single gen before your message. Did exactly like you said and great results (see my profile) - one image2video is 190 sec. (5070ti)
I finally found it was this lora that was giving me weird things like the guy sucking his own penis, or having a penis instead of a mouth and the girl giving a blowjob/kiss to it. It almost systematically happens with videos where the guy isn't visible at the start.
So be aware this can happen, it looks like the lora made the model understand blowjob to always involve kisses or something so it simply adds a dick to the guy's mouth.
Try removing “the man’s” or “1mans” and just describe the body parts if using I2V. It’s probably due to the captions being very specific about number and gender of participants and so something tells me it’s tangled up in some concepts/prompts. Omitting “man” might be beneficial.
Could you please make this for WAN 2.2 14b?
Got stuff on the way but have been pretty busy!
Didn't know I have a fetish/fantasy about having two dicks. The regular one, and the one that grows out of my left ball sack as the woman pulls her head back. Anyway - will work on my prompts to see if one girl can be working the penis and the other my balls. I do appreciate this LoRA.
extra penis, double penis has probably been in my negative for a long time lol
I also encountered the problem of men's faces. Later I found that the prompt words could not be "man", "man hip", or "man penis", but only the words hip and penis.
Yeah, that tracks. Much of our captions specify “the man’s penis” or whatever to try to avoid unwanted futa and other issues when it’s left ambiguous.
anyone know what pose is on the 1st picture? i never see a lora for that pose.
could someone share their workflow with the stacked loras? like wan\nsfw\dreamjob?
any change of a WAN 2.2 14B version of this masterpiece =) ?
The All in One version is nearly done and then this is next! It’s considerably slower to train and you’ve got to do two of them. :)
@c0ur4ge very cool! Will wait patiently =)
@c0ur4ge Eagerly waiting for your legendary works cant wait
Honestly, this works great on 2.2 with 1.0 strength on low noise and 2.0 on high
@Simon_Belmont well then you’re gonna really like what’s coming this weekend.
Go time.
@c0ur4ge ON IT!
whats up with this adding like a little bubble on the penis tip when doing a handjob? is there a way to fix that?
*******?
Sometimes it does happen. Just regenerate till you get a good seed. That's what I do.
add "cum, semen, precum" to the negatives, it can help
Details
Files
wan_dr34mj0b_t2v_HD.safetensors
Mirrors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan2.2_blow.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
T2V_wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
wan_dr34mj0b_t2v_HD.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.